3

Let's say I am using asynchronous streaming replication with the below configuration in a 3 node cluster with Postgres 10.4 and Patroni 1.4.4

    bootstrap:
      dcs:
        ttl: 30
        loop_wait: 10
        retry_timeout: 10
        maximum_lag_on_failover: 1048576
        postgresql:
          use_pg_rewind: true
          use_slots: true
          parameters:
            max_wal_senders: 10
            wal_keep_segments: 100
            max_replication_slots: 10

Let's assume that one of the replica nodes suddenly loses its connection to primary for a long time.

  1. In this case I think the size of WAL on the primary will keep on growing as it is not being consumed by the disconnected replica's replication slot. So is there any setting in patroni configuration which will remove the replica and remove its replication slot if it is disconnected from primary for x time duration?
  2. What is the recommended way to handle this case?
1

I would assume you are monitoring your DB cluster health, so a missing replica would pop up very soon. Also, it is a must to monitor disk space (running out of it might bring you into a situation that is not very easy to solve), so that would also catch this (later than sooner, usually).

Once you discover you have a replica that fell back, you have to investigate why it did so, and fix it - or remove the host from Patroni altogether. If under disk space pressure, remove the replication slot to free up WAL space. In a cloud setup, often simply terminating the host will solve all this by bringing up a new host. In any case, once you have a functioning host, you might have to reinit the Patroni node.

On the other hand, I'm afraid currently there is no mechanism for fencing off replicas that doesn't appear to come back (be it any actual implementation from removing the replication slot to anything more complex than that).

1

Replication Slots filling up disk?

Even if you specify use_slots in Patroni's configuration, the slot will be automatically removed if the replica fails to renew it's member key in the DCS (distributed consensus store). By default this means a slot for a disappeared replica will be dropped after 30 seconds (this corresponds to the ttl setting, which indicates how long leader and member keys will stay in the DCS) https://github.com/zalando/patroni/issues/1105

So unless you set up the slots yourself (with you could do with Patroni's permanent replication slot feature, though I wouldn't recommend that) your replica could very well fall behind infinitely. This means that there is no way for the replica to ever catch up, because the primary has already recycled some of the WAL that would be necessary to catch up.

How to avoid falling infinitely behind

The easy way to handle this scenario is to set wal_keep_segments high enough to cover your usual downtimes (network partitions, hardware maintenance etc.).

The recommended way to handle this is by implementing centralized WAL archiving, e.g. with pgBackRest, so that the lagging replica can retrieve missing WAL from the archive.

How to recover from falling infinitely behind

If your replica ever falls too far behind and the primary has already recycled some of the WAL that would be necessary to catch up, and you don't have a WAL archive, the replica needs to be recreated. This is very easily done running

patronictl -c /path/to/your/patroni.yaml reinit clustername membername

Depending on the create_replica_methods that you have defined, a backup will be used to "reimage" the replica. By default this is done through pg_basebackup.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.